This essay was submitted to the AI Progress Essay Contest, an initiative that focused on the timing and impact of transformative artificial intelligence. You can read the results of the contest and the winning essays here.

Alternative title: Artificial intelligence progress is bottlenecked on human intelligence progress.

Most people's predictions are epistemically vacuous

The cognitive process that happen when you ask someone to predict wether progress/a breaktrhough will happen in $magic_number years in any major topic, works usually the same way.

You ask someone a quantity/truth value and they will often try to determine it by pure intuition. For example, if you ask someone when will we find pharmacological drugs that significantly improve lifespan/healthspan they will assume the implicature of the loaded question that it has not already been achieved, despite in fact having been achieved multiple times.

Secondly, their gut feeling will tell that it is likely to happen in the next ~ 10-20 years.

If we assume that the person tested is a layman, he has actually close to zero understanding of what are the technical blockers/intricacies needed to actually achieve significant healthspan/lifespan progress. However he might have some confidence in its prediction magic number range. This is a strikingly potent example of a cognitive bias, this hipothethical person is not even realizing that his magic number is made out of an absent model, with zero predictive power. At best it might eventually use proxy variables that helps prediction indirectly as very weak heuristics.

Giving confidence to a value where we have abandonned the thinking process is often very pleasant and probably apply to most laymans predictions on complex, outside of their domain of erudition, topics.

The number range that non-erudite people generally give is not totally chosen at random though. It is an incouscious epistemic risk mitigation since 10-20 years is not immediate and therefore is not a high risk in topics where progress is "fast". What's funny though would be to plot the average time prediction range of people for complex topics over the years.

If you asked people in the 80s, 90s, 2000s, 10s, and 20s When will the first Artificial General Intelligence (AGI) be developped, my intuition tell me that at each of those decades, layman people would cyclycally tell approximately the same 10-20 years range. And thus the prediction cycle would continue for the whole century (until people get bored of AI winters eventually)

Because for the decade of the interrogated layman to matter, it would mean the tested subject has an altered predictive model, but for it to matter they would need to 1) actively follow the research advances, 2) the blockers towards the advancement of the AGI roadmap. This is something very few people actually do on Earth. Or 3) They would follow "big signals" on social media like the release of GPT-3, which does happen regularly.

Cool demos and the illusion of natural language understanding by the means of statistical tricks and bruteforcing (aka very large language models) has never been so perfected, reaching a point making us all want to believe. Of course the asymmetry that is the lack of signals regarging the absence of progress in key AGI and NLU blockers is striking and matrix the brain of regular AI enthusiasts. The absence of progress is a silent information that does not make buzz on social media. As such while my first thesis show that by default people guesstimate AGI for the next few decades, in a cyclical manner, this social media reporting assymetry biase people to says that AGI will happen sooner than later.

Scientific research is and will stay immature

One could say scientific research is in its infancy but a more accurate statement would be that scientific research is childish.

A common trope is the difference between the stereotypical scholar vs the engineer mindset, while the former has virtues (seeking accuracy, understanding) the former often pathologically fail at things like keeping track, focusing and funding the top priorities and maintaining pragmatism vs purity thinking.

Of course pragmatic and effective scholars do exists, however my experience (having read more than 10000 papers) has shown to me countless times that in many key topics, scholars systematically fails.

The field of medecine is the most potent one if you want to augment by a few order of magnitudes your misanthropy levels.

Medecine and more specifically pharmacology, AKA programming for the body has made multiple revolutionnary discoveries.

The root cause of most diseases and of death is the aging process AKA the self degradation of your body.

The aging process to be fully addressed has multiple factors at hand. However the main driving factor towards aging is oxidative stress. Breathing kill and more specifically Oxygen (hypoxia kills too though).

Oxidative products are unfortunately made as a byproduct of energy generation in the mitochondria. Those oxidative products liberate free radicals, which induce various damages randomly in the body (DNA mutations, conformational changes, apoptosis, etc)

Potent antioxidants like Gluthathione prodrugs (e.g. NAC) and SOD prodrugs (emoxypine) or atypical ones (e.g. echinochrome a) have many partially protecting effects in most diseases and increase healthspan and lifespan in mammals.

However classical antioxidants like NAC while useful and underprescribed, are not revolutionnary at all.

However scientists have realized than NAC dilutes in the cell plasma while most of the oxidative stress is generated inside the mitochondria.

Hence they designed mitochondria targeted antioxidants, that would be attired like a magnet inside the mitochondria.

SkQ1 was born and is 1000000 times more potent than NAC

While the effect of lifespan has conflicting results (up to a doubling according to some studies, a result consistant with the mitochondria targeted antioxidant fullerence C60), the effect on healthspan is certain and dramatic, such as having revolutionnary results for Alzheimer or even Parkinson. The algorithm is as follow for SkQ1, take a random disease, google SkQ1 + random disease on pubmed, observe generally revolutionnary results, profits.

Since oxidative stress is an almost universal cofactor amplifier of diseases and aging symptomatology, the possibilities are endless.

However this is old knowledge and those revolutionnary topics are systematically criminally underfunded. Sometimes by pure mediocrity (funding in mitochondria targeted papers), insane bureaucracy (clinical trials deaths induced by lack of clinical trials), or 3) by misaligned incentives.

Point 3) is major, the main most promising class of drugs in medecine are called peptides. For the most part they are endogenously produced and in general they cannot be patented. AKA a company cannot make profits with them hence this most promising chunk of medecine will mostly never become mainstream.

Covid has an obvious cure since day 1, it is more deadly in the elderly because elderly are immunosuppressed, more specifically their thymus gland self-atrophy (involution). By giving the peptide and thymic hormone Thymalin one can increase lymphocite T production by 610% hence making elderly immune system similar to a young one and therefore curing Covid.

I had this knowledge since day 1 and yet the scientific community except a few anticomformits humans on earth have consistently ignored it and this still applies to this days and will in fact apply for eternity. It is not a matter of when, when the scientific community systematically fails to identify even the most obvious solutions.

PNC-27 is one of the few revolutionnary treatments for cancer, unfortunately it is a peptide. I predict we will not see a clinical trial of it in the whole century (unless of course science stops being childish)

Those parallels I draw in medecine apply to AI research

Sure making papers reproducible (software) is much easier, AI does not suffer from bureaucratic rules too, however scholars consistently ignoring extremely potent, salient research do apply very well to AI.

One of the reason is that like in medecine, no one cares.

No one cares is really an essential insight to have, scholars might be paid to research but that does not protect them from conceptual attractors and repulsors.

Japaneses researchers have a study showing for the first time an effective treatment for the said incurable disease tinnitus. This was in the 90s and those people are Japanese (hence not in the pubmed attractor), hence no one knows this result, therefore no one will attempt to reproduce it scientifically. It's not that the information is that hard to find per se, its just that its difficulty is above the treshold of caring that researchers have.

Before anything, humans want to reinvent the wheel themselves, siloed research is a cancer that affects everyone to a significant extent.

Secondly, there is no issue tracker for tracking the progress on specific research issues.

This alone is a revolutionary statement, if you have experience the beauty of an issue tracker like Github offer for software developers and collaborative editing a la wikipedia, and imagime a mix of both, then you could see what I see, the biggest missed opportunity in making science progress traçable in real time, accelerating the identification of blockers, exhaustiveness of the solutions explored on the mental search space and making people from transverse expertise meet and complement each others.

Instead of this, people simply publish their new siloed NIH paper in the ocean of NIH papers.

The pace of AI progress

Paperswithcode is the reference source of truth for tracking AI progress in a given task.

I have read leaderboards in all NLP tasks since 2019 and I can assert approximately that on most tasks, progress is on the order of ~1% accuracy gain per year, which is slow and fast at the same time.

An excellent idea that no-one has done to my knowledge regarding a quantitative analysis of AI progress would therefore be to do a global average of accuracy gains per year for all NLP and Vision tasks, it would make an interesting plot (although with coefficients for lowering the impact of low importance tasks)

While OpenAI and social media like seeing progress on the task of predicting the next most likely word in a sentence (language modeling), giving the statistical illusion of true causal NLU, the real NLP tasks tell a different story of progress.

There has been major advances in key tasks such as dependency parsing however many major NLP tasks have progress stalled since 2019.

There are some researchers that cares about understanding language though, you will generally found them in an essential field for AI research that ~no-one study: Computational linguistics

AGI has no established roadmap

A true NLU system needs to combine ~50 NLP SOTA tasks together to even begin to do proper semantic parsing. Now keeps in mind that error between those models adds-up or even sometimes multiplies. Since most tasks have <95% accuracy and many at <90% accuracy, this means that at best we can only begin to build proof of concepts of a true natural language understanding system, which would only works for a short time until it accumulates too many errors.

Almost each of those tasks are necessary and for many tasks, no one is actively working on them.

If the 1% of progress were sustained well NLP would reach 100% accuracy in 10-20 years in most key tasks however, the progress is already hitting an inevitable diminishing return of low hanging fruits and there are many signals of an incoming plateau/ai winter.

Quantifier scope disambiguation is an example of the many necessary, ignored tasks/

"A doctor lives in every city."

We have underlined the two quantifiers in the sentence: “A” and “every”. The former is an existential quantifier and the latter is a universal quantifier. There are two possible interpretations

The first is that there exists a doctor who resides in every city. The second and more plausible interpretation of the example sentence is that for every city, there exists a doctor who lives there (not necessarily the same doctor).

This foundational tasks has had like what, 3-5 researchers allocated to it in the last 2 decades ? At some point it's not a matter of decades but of human resources.

Researchers is QSD will explain that language models cannot be applied to it since no-one paid for building a large labeled dataset of QSD. I have seen it many times in NLP, we lack essential datasets, funding is not allocated to the right place and until those issues are solved AGI will not come tomorrow but never.

Yes, AGI will never come if the blockers are consistently left unaddressed, which is what's happening since the 2000s.

It is trivial to consistently improve the state of the art in any NLP task

Yes, you read this right, anyone can trivially give a few % accuracy increase in any NLP tasks.

Here's how:

1) Identify that many researchers have found effective, generally applicable improvements

2) observe they are consistently ignored by researchers (because of NIH syndrome) AND that they can be synergetically combined

3) profits.

Let's say you wanna be the leader in coreference resolution for example ?

Well be the first to use XLnet (or deberta v3) as a much superior baseline vs BERT. ~+3% accuracy gains

Optionally, use the activation function mish +~0.5-1% gain

optionally use RAdam + lookahead + ~1-2%

bonus: gradient centralization, etc

Each of those is sufficent to beat the SOTA in ~any task and only one human has combined them unfortunately untested for NLP as of the time of writing since he is involved in computer vision.

I once told Baidu to leverage those, they said it was a promising research direction, and those trivial yet major optimizations have of course been ignored even when pointed in front of their eyes.

I am the one that told researchers in the key task of dependency parsing to leverage XLnet, and for once they listened and gained better accuracy.

Conclusion$true$

There is a deficit in fundamentally original ideas in NLP.

LSTM and RNN are very old ideas. While there might (or not) be genius in how to achieve a transformer implementation, if we dezoom a little, intuitively the idea (not the technical how) is not very original. All the optimizations I mentionned above are not that original and essentially are only incremental optimizations subject to diminishing returns, not a paradigm shift.

If we had many original and potent ideas and it was a matter of finding the how then I could believe AGI or at least true NLU could come in the next 15-20 years, however what is understated is that we have a shortage of sound, high levels ideas in how to achieve AGI, we are not even blocked on the technical details of the how, we are blocked on the what.

If you wanna get more intuition as to why neural networks are just a small piece in the equation, I invite you to study the Abstract Reasoning Challenge.

This is one of the most interesting benchmark in AI and the solutions in the leaderboards are either expert systems, meta-rules to make rules or attempts at the ambitious (but exponential) goal of automated program synthesis. Neural networks are essentially unfit for this task, even though very large language model become bearable at zero shot or few shot learning, their accuracy on zero shot is still essentially mediocre.

I had hopes that key insights would be found by studying the worm C.elegans, which only has ~300 neurons and has complex behaviours, that is, until I realized once again no-one cares and that there is zero funding despite being key to unblocking the incoming AGI winter.

To conclude, for the building of the first AGI, we firstly need to signiticantly restructure the way we make science and the way we allocate cognitive and monetary resources.

In other words, the advance towards an artificial intelligence might require something close to a collaborative human superintelligence.